Ancient Books Meet Engines: When a 300-Year-Old Verdict Faces the Machine

For centuries, a verdict in xiangqi was true for one reason: it was printed in a book, and generations of players passed that book hand to hand. The system worked astonishingly well — but it had one fatal weakness: errors got printed too, and a printed error was then copied faithfully across books and centuries. We met the phenomenon head-on while building the Xiangqi Position Dictionary: there are positions where both of our independent transcription sources print the identical diagram twice with two different verdicts — a draw in one place, a Red win in another. Nobody can say which is right; we held that position back and published nothing. This article describes how modern engines are re-judging those classical verdicts, through the dictionary's own verification queue — which so far holds 272 confirmed positions, and exactly one overturned.

Two Swallows Fighting in Flight (Mengru Shenji, game 112): the book says mate in 11 — the engine confirms it to the move.

Our protocol: what counts as "verified"?

The dictionary uses three labels, and the borders between them are strict:

• ✓ Verified: an engine (Fairy-Stockfish with an NNUE evaluation network — the same open-source engine family as Pikafish, which now sets the standard for xiangqi) finds a complete forced win, and that line is replayed and checked for legality move by move. For a position the book calls a draw, an engine-found forced win for either side is likewise an overturn.
• ⚖️ Disputed: the engine finds a forced win against the book's verdict.
• Unverified — result as recorded in the book: everything else, including positions an engine has examined without reaching a conclusion. A cut-short search is not a verdict.

That last principle sounds obvious, but it is the one that saves you from the engine era's most common mistake: overclaiming.

Four cases from our own verification queue

The most beautiful confirmation: Two Swallows Fighting in Flight. This position from the Mengru Shenji manual is recorded as mate in 11 for Red. On verification, the book's main line replayed legally through all 21 plies, ending in exactly the mate on move 11 — and an independent source dataset also records the Pikafish engine confirming the same number, 11. The detail matters: all eleven Red moves in the solution are checks, the two horses handing an unbroken chain of checks back and forth. The ancients had no machines; they found a perfect checking sequence by hand and wrote it down without a single error. Here the engine does not diminish the old book — it certifies the craftsmanship of its authors.

The "undecided" cases that decided: Zhuge Leaves the Cottage and Destroying the Four Bandits. Both positions sat in our "searched but undecided" bucket, because earlier engine runs had been cut off mid-search — the lines were too long, the time too short. Given full runs, both closed: Zhuge Leaves the Cottage is mate in 22, Destroying the Four Bandits is mate in 26, each with a complete, legal main line. The lesson generalises to anyone using engines to judge old positions: a truncated search produces meaningless numbers, and those meaningless numbers nearly became our "verification results".

The label we refused to flip: No Peace at Home. The book records a very long winning line for this position. But the transmitted line is truncated — it does not reach mate. Our engine has not closed the case, and the book's line cannot be fully replay-verified. So the entry stays unverified, even though every sign suggests the book is right. A dictionary should rather be slow with a label than wrong with one.

The single overturn: Wild Horse Ploughing the Field. Among more than 1,500 entries, exactly one currently carries the disputed label: Wild Horse Ploughing the Field, one of the Four Greats. Its original 1879 form records a draw; engine-era research leans toward a Black win. A verdict printed in the nineteenth century and copied through at least four classical manuals did not survive. This is the case we cite most often, because it proves the label system has teeth: "disputed" is not decoration.

What changes when the machine referees?

First, authority moves from print to proof. Once, arguing with a book required being a master; today any club player with a strong engine can ask the question — and if they genuinely find a refutation, the question is worth the same.

Second, heritage does not lose value when a verdict falls. The main line of an overturned position is still chess literature: it shows what the ancients thought, how far they calculated, and where they almost had it. That is why our dictionary never deletes a book verdict when an engine overturns it — we place the two side by side, each clearly attributed.

Third — and this part is deliberate design — verification should not belong to our servers alone. In the dictionary, any member can take a position against the engine at its strongest level: match the book's recorded result and you are credited as a solver of that position — a confirmation vote cast by actually playing; but hold the defending side to a better result than the book records, and your game becomes refutation evidence, the entry jumps to the front of the engine queue, and if the engine confirms, your name is recorded permanently on the entry's page as its breaker. A heritage guarded by a community should also be verified by that community — the engine is only the final referee, not the only player.

The dictionary currently stands at 272 verified entries out of more than 1,500 — a modest number, kept modest on purpose. Everything else plainly says "result as recorded in the book": not a promise, but an invitation to verify.

From the SynerWorlds series on classic xiangqi positions: The Four Greats · Jianghu traps and the secret manuals · Seven Stars Gathering, dissected.